Problem
In normal operation the controller should track the path, remain inside the road boundaries, respect actuator limits, and avoid obstacles. My research adds a second regime: if the remaining state and actuator limits make avoidance infeasible, the optimizer should choose the remaining action that reduces collision severity.
MuJoCo validation platform
I built a forward Ackermann vehicle environment with steering geometry, tire-road friction, actuator constraints, track boundaries, obstacle-avoidance scenarios, and instrumented collision tests. The head-on collision harness logs bumper force, impulse, velocity, acceleration, momentum, and trajectory data while allowing approach speed and obstacle mass ratio to be varied.
Control development
The obstacle-avoidance testbed progressed from a lightweight lateral controller to an NMPC formulation using the state [x, y, ψ, v] and drive/steering commands. The controller predicts over a finite horizon, follows smooth lane-change references, respects actuator limits, penalizes predicted road-edge violations, and applies only the first optimized command before replanning.
Why computation time matters
A theoretically better trajectory is not useful if computing it consumes the remaining reaction time. The project therefore studies the tradeoff between horizon length, model fidelity, replanning frequency, and the amount of time available before impact.
Current research questions
Current work focuses on selecting a physically meaningful collision-severity surrogate, incorporating tire-force and friction limits, and designing a controller architecture in which a future damage metric can be added without rebuilding the path-tracking and avoidance stack.
Read the research overview →